Continuous Preference

Trend Mining for Optimal

Jungmok Ma
Enterprise Systems Optimization Laboratory,
Department of Industrial and
Enterprise Systems Engineering,
University of Illinois at Urbana-Champaign,
Urbana, IL 61801
e-mail: jma15@illinois.edu
Harrison M. Kim1
Associate Professor
Enterprise Systems Optimization Laboratory,
Department of Industrial and Enterprise
Systems Engineering,
University of Illinois at Urbana-Champaign,
Urbana, IL 61801
e-mail: hmkim@illinois.edu
Product Design With
Multiple Profit Cycles
Product and design analytics is emerging as a promising area for the analysis of large-
scale data and usage of the extracted knowledge for the design of optimal system. The
continuous preference trend mining (CPTM) algorithm and application proposed in this
study address some fundamental challenges in the context of product and design ana-
lytics. The first contribution is the development of a new predictive trend mining tech-
nique that captures a hidden trend of customer purchase patterns from accumulated
transactional data. Unlike traditional, static data mining algorithms, the CPTM does not
assume stationarity but dynamically extracts valuable knowledge from customers over
time. By generating trend embedded future data, the CPTM algorithm not only shows
higher prediction accuracy in comparison with well-known static models but also pro-
vides essential properties that could not be achieved with previously proposed models:
utilizing historical data selectively, avoiding an over-fitting problem, identifying perform-
ance information of a constructed model, and allowing a numeric prediction. The second
contribution is the formulation of the initial design problem which can reveal an opportu-
nity for multiple profit cycles. This mathematical formulation enables design engineers to
optimize product design over multiple life cycles while reflecting customer preferences
and technological obsolescence using the CPTM algorithm. For illustration, the devel-
oped framework is applied to an example of tablet PC design in leasing market and the
result shows that the determination of optimal design is achieved over multiple life cycles.
[DOI: 10.1115/1.4026937]

Keywords: continuous preference trend mining, product and design analytics, design for
multiple profit cycles, model tree

1 Introduction and Background Protection Agency (EPA) estimated that 2.37  106 short tons of
electronic products were ready for the end-of-life processes in
Product and design analytics relates and utilizes large-scale
2009 [6]. Compared to the amount in 1999, almost 50% have
data for design decisions throughout the life cycle of complex
increased and only 25% of them were gathered for recycling. The
products. The area of analytics is emerging as a promising area
second reason is the fact that electronic products contain toxic and
for the use of large-scale data generated by the users, original
hazardous materials [7]. Lead, mercury, nickel, and palladium are
equipment manufacturers (OEMs), markets, and the public, which
examples that present negative environmental and human health
are often freely available. Recent progress in this area has gar-
effects. Reckless landfills are not an optimal solution. A third fac-
nered notable achievements in predictive product design, for
tor of the problem of recovering these products is that electronic
example, merging predictive data mining and knowledge discov-
products also contain reusable and valuable resources, such as
ery methods with product portfolio design [1], predictive trend
gold, copper, tin, nickel, etc. [7]. Efficient and systematic methods
mining for product portfolio design [2], defining design analytics
to recover the reusable parts and resources are needed. Fourth,
[3], and linking on-line reviews with product attributes [4,5], to
more regulations and responsibilities are emerging. The countries
name a few. The area of product and design analytics could be
in the European Union have already begun adopting product take-
particularly beneficial for the area of product life cycle design and
back policy (Extended Producer Responsibility, EPR) since 1991
recovery, as product end-of-use state (e.g., life cycle length, prod-
[8]. The U.S. has also introduced more EPR laws recently com-
uct condition, product recovery decision making) is hard to pre-
pared to 2006 [9]. Fifth, “green consumers” [10] give more pres-
dict. The current paper shows that product and design analytics
sure to companies regarding their “green” image. Now, their
can serve as a foundation to link product prelife (design and man-
increased awareness of sustainability is a critical factor in deter-
ufacturing), usage, and end-of-life operations with the develop-
mining the demand of target products. Lastly, product recovery
ment of the trend mining algorithm.
and recycling are known to reduce the necessity for fuel consump-
tion and landfill space, and provide substantial benefits environ-
1.1 Recovery of End-Of-Life Electronics. Recovery of end- mentally [11].
of-life electronic products has become an urgent problem that Some OEMs such as Caterpillar, Xerox, and Sony have shown
requires design engineers’ attention due to multiple reasons. The first that a proper recovery system of their end-of-life products not
reason is the fast growing e-waste stream. The U.S. Environmental only extends their products’ lives and gives some environmental
benefits, but also allows for multiple profit cycles [12–14]. These
1
Corresponding author. OEMs consider the end-of-life stage as the “relife” stage
Contributed by the Design Automation Committee of ASME for publication in
the JOURNAL OF MECHANICAL DESIGN. Manuscript received May 23, 2013; final
and return take-back components to “same-as-new” condition
manuscript received February 14, 2014; published online April 11, 2014. Assoc. to customers. The relife processes or recovery options include
Editor: Bernard Yannou. reuse, repair, refurbishment, cannibalization, and recycling. By

Journal of Mechanical Design Copyright V

C 2014 by ASME JUNE 2014, Vol. 136 / 061002-1

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 introducing remanufactured products back to the market, compa-
nies can find new profit opportunities and establish a greener or
more environmentally friendly image.
The evidence from these OEMs indicates that the construction
of a system for recovery can be a hidden source of profit. How-
ever, many factors should be considered in order to determine the
profitability of the recovery system. Possible sources of uncertain-
ties are the product’s life, the state (product condition) after its
life, available quantities for recovery, the reliability of a remanu-
factured product, customer preferences, and technological obso-
lescence. Product and design analytics can provide a systematic
way to explain these uncertainties for multiple recoveries or mul-
tiple life cycles.
1.2 Predictive Trend Mining for Product and Design
Analytics. Data mining in the context of product and design ana-
lytics was suggested as an alternative for knowledge extraction
[15]. Traditionally, there are a few methods for capturing cus-
tomer requirements and preferences such as quality function
deployment, conjoint analysis, and discrete choice analysis. These
methods resort to direct or close interactions with target customers
and generate stated preference data. The strength of using data
mining models is to utilize revealed preference data or accumu-
lated data sets related to customers’ actual behavior (e.g., transac-
tional data, sales, and on-line reviews) that usually have
characteristics of large volume, unstructured form, and timeliness.
Predictive trend mining is a new and emerging data mining
area, which is also known as change mining [16,17] or learning
concept drift [18]. Unlike traditional static data mining models
with the assumption of stationarity, the predictive trend mining is
a dynamic and adaptive model that captures trend or change of
customer preferences over time.
A tree based data mining algorithm with predictability was
used in the predictive trend mining. Tucker and Kim [2] proposed
the Discrete Preference Trend Mining (DPTM) algorithm and
suggested a classification of attributes as standard, nonstandard
and obsolete with respect to a class variable for guiding design
engineers. The attributes or features are also known as independ-
ent and explanatory variables, and the class variable is a depend-
ent and response variable. It should be noted that due to the fact
that the algorithm was developed to deal with discrete class varia-
bles and attributes for product portfolio concept generation, the
term Discrete is added to the original name PTM. For example,
five discrete prices {$99, $149, $179, $199, $249} were used as
the class variable and no design problem was provided. Ma et al.
[19] extended the work and proposed that the predictive trend
mining technique called demand trend mining (DTM) can benefit
optimal life cycle design problems. Utility was used as the dis-
crete class variable and discrete choice analysis was utilized to
calculate expected market shares. However, the nature of optimal
design problems often requires continuous variables, e.g., price,
cost, demand, etc., and discrete class variables might limit the
application of design problems. In order to allow continuous vari-
ables while capturing a trend, a new method, Continuous Prefer-
ence Trend Mining (CPTM), is presented in this paper.
Support vector machine (SVM) [20] is another data mining tool
that can be used in the predictive trend mining. The SVM learns
by example to classify different groups. Klinkenberg [18] discussed
several methods to handle concept drifts based on the SVM. With
concept drifts, different weighting schemes for historical data are
possible, i.e., each data point over an extended period of time can
be removed or utilized based on its age by allocating individual
weights. Klinkenberg showed that the performance of his adaptive
techniques outperformed that of simple heuristic approaches such
as using all data or the most recent data in his simulated experi-
ments. However, the proposed SVM based techniques are not for
numeric prediction but for simple binary classification.
The CPTM algorithm will shed light on the initial design prob-
lem which has an opportunity for multiple profit cycles. If there
061002-2 / Vol. 136, JUNE 2014
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms
are multiple recovery chances for end-of-life electronics in the
near future, a trend of customer preferences and technological ob-
solescence will be traced and captured at the target time for the
optimal initial design. The captured information will then be
merged with a product design problem.
1.3 Merging Predictive Trend Mining With Design for
Multiple Life Cycles. Design for multiple life cycles is a design
paradigm that enables design engineers to close the loop of a
product life cycle and to manage its multiple life cycles. Leasing
or sales of service is a representative example of the management
of multiple life cycles as shown in Fig. 1. After designing and
manufacturing a product, a lessor (a person possessing a product
that is being leased) would lease the product to a lessee (user of
the product being leased). At the end of the lease contract or the
usage stage, the lessor would take back the product and determine
a proper recovery option. If it is profitable, the lessor would lease
the product again for a multiple periods of time. Eventually, a
product would generate multiple profit cycles, k. Many studies
showed that the initial design of a product would determine
70–85% of total life cycle cost and environmental impact
[21–23], so the selection of initial design attributes is the focus in
this paper, especially from the economic perspective.
In order to combine the design problem with the CPTM, design
for multiple life cycles is proposed to be formulated as an optimi-
zation problem. The formulation determines the design attributes
that maximize the total life cycle profit and generate multiple
profit cycles. Only a few studies [24,19] provided mathematical
models that realize the total profit from both prelife (design and
manufacturing) and end-of-life stages. Design for multiple life
cycles will extend these studies.
1.4 Research Approach. In this paper, the CPTM algorithm
is developed in order to take large sets of transactional data and
extract valuable knowledge of customer purchase patterns. The
CPTM is then tested with generated and real data sets. The archi-
tecture of the CPTM will help to predict the target class variable
that reflects trend of customer preferences and technological obso-
lescence over time. By merging the continuous, predictive trend
mining technique with an optimization model, the proposed
framework will produce an optimal product design that maximizes
a total unit profit and eventually reveals an opportunity for multi-
ple profit cycles.
The rest of the paper is organized as follows. In Sec. 2, the
entire methodology is explored with the CPTM algorithm and an
optimal product design model for multiple profit cycles. Section 3
presents performance tests of the CPTM with various data sets.
An illustration example of tablet PC design is provided in Sec. 4,
Fig. 1 Product life cycle in leasing market
Transactions of the ASME
 where X is a set of attributes, T is a data set, and Tj is a subset of
the data T after splitting. The denominator represents the informa-
tion generated by splitting the data set T into n partitions. The nu-
merator represents the amount of uncertainty reduction by
splitting on attribute x. Entropy quantifies the expected value of
the information in bits. p(ci) represents the probability mass func-
tion of a class variable ci and k is the number of class values.
The concept of building a tree model based on the predicted
Gain Ratio was initially proposed by B€ottcher and Spott [16]. The
predicted Gain Ratio provides a way to build a future classifica-
tion tree without real data but there are some disadvantages. First,
there is a strong possibility of over-fitting since no pruning pro-
cess is suggested. Highly branching trees risk over-fitting the
Fig. 2 Overall flow of methodology training data and performing poorly on new samples. Pruning can
help to determine the optimal size of trees. Second, no perform-
and the conclusion and future research directions are presented in
ance result of built models can be estimated since there is no test
Sec. 5.
data. Third, the Gain Ratio based methods are only applicable to
classification models or discrete class variables. It will be shown
2 Methodology that by generating the target data, the CPTM algorithm can
The entire framework is divided into two phases. Phase 1 is to provide a way to utilize historical data selectively, avoid an
implement the CPTM algorithm, which entails data preprocessing, over-fitting problem, and identify performance information of
trend embedded future data generation, and model tree induction constructed model. More importantly, the CPTM adopts a tree
as shown in Fig. 2. Phase 2 involves an optimal product design for induction model that allows use of continuous class variables.
multiple profit cycles by combining the predictive model built Usually the process of data mining consists of data collection
from the CPTM. and selection, cleaning and transformation, pattern discovery, and
The schematic of the CPTM algorithm shown in Fig. 3 con- interpretation. In the product design domain, text and web mining
structs a predictive model (model tree in Sec. 2.1.3) at time n þ h [28] provides a way for design engineers to collect and analyze
or h periods ahead based on the historical data sets from time 1 to customer preference data (e.g., review data), including identifying
n. The core part of the algorithm is the generation of trend embed- product attributes and modeling customer ratings [29–33]. In this
ded data. Geometric sampling is developed to capture the trend of paper, our focus is limited to the pattern discovery and interpreta-
the relationship between design attributes and class variables by tion stage.
sampling normalized historical data selectively (i.e., ‹ and ›).
Automatic time series forecasting proposed by Hyndman et al. 2.1 Phase 1: Continuous Preference Trend Mining
[25] is used to predict future values of design attributes and class
variables. By applying the predicted values to the normalized 2.1.1 Data Preprocessing. The first step, data preprocessing
sampled data, unseen future data at time n þ h or Dnþh can be gen- is a data preparation technique for trend mining. It starts by gath-
erated (i.e., fi). Finally, the future model tree or MTnþh can be ering and organizing n-time stamped transactional data sets. An
built based on the trend embedded data (i.e., fl). The two dotted example of a data set is shown in Table 6. The data set consists of
boxes represent the predicted data and model, which are not avail- a set of attributes and one class variable. In the example, there are
able initially. eight different attributes of a product and a class variable, price,
The DPTM, on the other hand, builds a predictive model (deci- which customers paid in their transactions. Even though any class
sion tree [26]) based on predicted values of Gain Ratio, one of variables that researchers are interested in can be selected, paid
splitting measures. The mathematical form of the Gain Ratio is price or market value was used in this study since it is directly
defined as [27,2] related to customer preferences. Sales or demand can be another
candidate. There is no restriction on the data except that the class
EntropyðTÞ  Entropyx ðTÞ variable should be continuous. Both discrete and continuous
Gain RatioðXÞ ¼ Xn attributes can be dealt with by using the proposed approach. In
jTj j jTj j
  log2 this paper, only one class variable is modeled. In order to allow
j¼1
jTj jTj more than one class variable, a multivariate tree [34] can be used
X
k X
n instead of a univariate tree (i.e., model tree).
jTj j Next, discrete attributes are transformed to a set of binary
 pðci Þ  log2 pðci Þ½bits   EntropyðTj Þ
i¼1 j¼1
jTj attributes. In the case of attributes with significant improvement
¼ X
n in their values (which are component based attributes, e.g. Hard
jTj j jTj j
  log2 drive, CPU, etc.), the values need to be expressed as a genera-
j¼1
jTj jTj tional difference [35]. The generational difference is a relative
(1) scale that can be acquired by comparing the generational gap
between the target part and the latest cutting-edge part which cor-
responds to the minimum generation or zero. As time passes, a
new part is introduced in the market and the generational differ-
ence of the existing part is increased. We assume that customers
perceive the relative generational gap of components with a given
time, and a company has expected values or a roadmap of the
components in the near future. The generational difference is uti-
lized to represent the technological obsolescence and its effect
over time. In the appendix, an example of generational difference
is shown over time and the cutting-edge part has a value of zero.
Before moving to the next step, it can be checked whether
structural changes or trends are in the data. In this paper, two pos-
sible trends are identified. First, levels under each attribute and
Fig. 3 A schematic of CPTM Algorithm class variable can have increasing, decreasing or cyclical patterns.

Journal of Mechanical Design JUNE 2014, Vol. 136 / 061002-3

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 For example, the display size of cell phones can have an increas-
ing pattern. In order to detect this kind of trend, it is useful to visu-
alize data. Statistically, Spearman’s rho test of trend and Mann-
Kendalls tau test of trend are available [36]. Second, there are
some trends in terms of relationship between design attributes and
class variables over time. For example, the memory size of note-
books might be an important factor for the purchase of the prod-
ucts a couple of years ago but some technological advances can
change the importance of the memory size in the next year. There
is no known method to detect this kind of trend but one possible
way is to apply the tests of trend to the coefficients of regression
models. If both trends are not detected (i.e., static case), the
CPTM will generate the same result with the simple model tree,
and other static models (e.g., regression, neural network, SVM,
etc.) can be applied to the latest data set or the entire data set
depending on the characteristics of data.
2.1.2 Trend Embedded Future Data Generation. The second
step is the generation of trend embedded future data. In the previ-
ous section, two different trends were introduced in data. The
automatic time series forecasting is the technique that captures the
first type of trend, and the geometric sampling that is newly pro-
posed in this paper helps to capture an underlying relationship
between design attributes and a class variable (i.e., the second
type of trend) by selectively utilizing historical data.
When there are a series of time stamped data points,
{x1 ; x2 ; …; xn }, where xt stands for a data point at time t, a couple
of different techniques can be applied to forecast a data point at
n þ 1 or one-step-ahead forecast. As a heuristic, it is possible to
take either the latest data point or the average of all historical data
for the forecast. Simple moving average is a method to smooth a
time series over last k observations though the selection of k can
be a heuristic. Exponential smoothing is one of well-known time
series analysis methods, and the simplest form is given by [25]
xn ¼ kxn þ kð1  kÞxn1 þ kð1  kÞ2 xn2 þ   
x^nþ1 ¼ kxn þ ð1  kÞ^
þ kð1  kÞ3 xn3 þ kð1  kÞn1 x1 þ ð1  kÞn x^1 (2)
where x^t is a forecast at time t and k is a constant between 0 and
1. The exponential smoothing is a weighted moving average of all
time series with exponentially decreasing weights defined by k.
The expanded form shows that recent values have a greater weight
than old ones. A total of 30 exponential smoothing models are
classified based on the combination of trend, seasonal, and error
components. There are two error components (additive and
multiplicative), three seasonal components (none, additive, and
multiplicative), and five trend components (none, additive, multi-
plicative, additive-damped, and multiplicative-damped). For
example, a model with all additive components can be expressed
as (trend þ seasonal þ error) and a model with all multiplicative
components is (trend  seasonal  error). Hyndman et al. [25]
provided all the classifications. We adopted the automatic fore-
casting method [37]. First, apply all the 30 exponential smoothing
models and estimate initial states and parameters using maximum
likelihood estimation. Second, choose the best model according to
one of the following criteria: Akaike’s information criterion
(AIC), corrected Akaike’s information criterion (AICc) or Bayes-
ian information criterion (BIC) [37].
After the geometric sampling process which will be introduced
shortly, the sampled normalized data set for the target time is
finally transformed to the real value data by applying predicted
values of each attribute and class variable. The minimum and
maximum values of each attribute and class at the target time are
predicted (i.e., the first type of trend) by the automatic time series
forecasting algorithm. By adopting the automatic algorithm, users
do not need to resort to their own knowledge for models and
parameters.
The second type of trend cannot be captured by time series
analysis methods since the underlying relationship between design
061002-4 / Vol. 136, JUNE 2014
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms
attributes and class variables is hidden. However, similar to the
exponential smoothing, required traits include dynamically utiliz-
ing all past observations and applying decreasing weights in order
to reflect underlying trends of the relationship between design
attributes and class variables. Previously proposed trend mining
models [16,2] did not consider the dynamics of relative impor-
tance of historical data. For example, B€ottcher and Spott [16] used
a polynomial regression method to predict the future Gain Ratio.
This implicitly gave equal weights for all historical data. If histor-
ical data is available and older data sets contain more errors (this
can be viewed as outliers), the accuracy of the predictive model
will be diminished. The CPTM algorithm, on the other hand, pro-
vides a dynamic selection of historical data for the reflection of
upcoming hidden trends by assigning exponentially decreasing
weights to old data sets.
The geometric sampling is a method to sample historical data
selectively for the second type of trend. Before sampling, each at-
tribute and class variable should be normalized within a single
time step. The tth term of the geometric sampling or at which
gives the number of instances (data points) that needs to be
sampled at time t is given by
at ¼ að1  aÞnt (3)
where a is an original number of instances, (1-a) is a common ra-
tio in geometric series, a is a smoothing factor ð0  a  1Þ and n
is the latest time. The smoothing factor a can be considered a
characteristic of product domain in terms of relationship between
design attributes and class variable. Table 1 indicates that when a
is close to 1, only the latest data set is useful, and the product do-
main is technology sensitive and rapidly changing. When a is
close to 0, all data sets are valid for future target time and the
product domain has a quite insensitive and slowly changing
characteristic.
In the geometric sampling, a is defined as a smoothing factor to
generate t ¼ n data using t ¼ 1 to t ¼ n– 1 data when t ¼ 1 to t ¼ n
data are available. a is obtained by
arg min E (4)
a
where E is a performance measure (e.g., error metrics such as
mean absolute error, root mean-squared error, and relative squared
error, etc.) tested for a model tree constructed from t ¼ 1 to
t ¼ n  1 data sets (as train data) with t ¼ n data (as test data).
The data for building a model tree can be sampled by Eq. (3). A
model tree will be introduced in the next section.
For example, if t ¼ 1 to t ¼ 10 normalized data sets are avail-
able, using t ¼ 1 to t ¼ 9 data sets, a model tree can be constructed
with different a values and predicted values of attributes and class
variables at t ¼ 10, and validated with t ¼ 10 data. Table 2 shows
the best a example in terms of the performance measure, mean
absolute error which is the average deviation between predicted
and observed class variable price, with simulated data sets (each
has a thousand instances or a ¼ 1000). For a ¼ 0.9, the number of
total instances (1111) comes from a thousand instances
(1000ð1  0:9Þ0 ) from t ¼ 9 data, a hundred instances
(1000ð1  0:9Þ1 ) from t ¼ 8 data, ten instances (1000ð1  0:9Þ2 )
from t ¼ 7 data, and one instance (1000ð1  0:9Þ3 ) from t ¼ 6 data
based on Eq. (3). The required numbers are sampled randomly
using a random number generator. The sampled normalized data
becomes real value data after applying predicted values of
Table 1 a value and product domain
a value Sampling Product domain
1 only the latest data set technology sensitive, drastically changing
0 all data sets insensitive and slowly changing
Transactions of the ASME
 Table 2 Example of best a selection

a 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Mean absolute error (MAE) 39.6 37.4 35.4 34.8 32.8 33.0 33.4 33.7 34.7 35.5 37.0
Total number of instances 9000 6125 4321 3199 2477 1998 1666 1427 1250 1111 1000

Fig. 4 Graphical example of trend embedded data generation

attributes and class variables. Then, a model tree can be built models classify class values without data so that no comparison
based on this data and tested with t ¼ 10 data. The best a is 0.4 can be made between a node and subtree for the pruning. In
based on the performance measure, mean absolute error. Sec. 3, pruning in the model tree algorithm will be introduced.
Based on the selected a, a number of required instances for
each data set is determined and sampled as in the example but
using t ¼ 1 to t ¼ 10 data sets this time. By applying predicted val- 2.1.3 Model Tree Induction. The third step is to build a model
ues of attributes and class variables at a target time to the sampled based on the newly generated data set from the second step. In
normalized data, trend embedded future data is finally generated this step, the knowledge and hidden patterns between the new val-
at the target time. Table 2 shows that the total number of samples ues of attributes and class variables are mined using a model tree.
can be varied depending on the selected a. Based on the smooth- The result of the model tree is a piecewise linear regression equa-
ing factor, only the latest data or all data can be used in the tion depending on a given data set, which can approximate nonlin-
extreme case. ear functions. Figure 5 shows an example of a model tree. The
A graphical example of the trend embedded future data genera- model tree gives three different linear models to express the nonli-
tion is depicted in Fig. 4. The values of original data sets are nor- nearity with two attributes: A and B. On the other hand, a decision
malized within a single time step and then sampled using the tree that was used in the other trend mining algorithms classifies
geometric sampling method with the selected a. In the example, discrete or categorical class variables.
suppose that the first and the last instances were sampled from the The M5 model tree was initially proposed by Quinlan [38].
t ¼ 10 data set. By applying predicted minimum and maximum After comprehensive descriptions of model tree induction includ-
values from the time series prediction technique, real values are ing a pseudocode by Wang and Witten [39], the model tree has
predicted at the target time. For example, the display size is get-
ting bigger, and the generated target data set reflects the trend
(e.g., refer to the small arrows).
By generating future data, two advantages can be achieved.
First, performance information of built models can be provided
similar to normal data mining processes. The predicted Gain Ratio
based models in Sec. 1.2 cannot give test data but the generated
data from the CPTM can work as test and validation data. The 10-
fold cross-validation technique [28] is a popular way to get the
performance (i.e., prediction of class variables) information when
a validation data is not available. In the 10-fold cross-validation,
the generated data are randomly partitioned into 10 subsamples
and validation processes are repeated 10 times. Each time a model
is built using 9 subsamples and validated with one remaining sub-
sample. Then, an average performance error can be estimated.
Second, pruning can be implemented based on the generated data
to reduce the risk of over-fitting. The predicted Gain Ratio based Fig. 5 Example of model tree

Journal of Mechanical Design JUNE 2014, Vol. 136 / 061002-5

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 received attention from researchers. Wang and Witten’s model
tree algorithm is known as M5P. The basic operation is splitting,
and the splitting is based on standard deviation reduction (SDR)
in following equation [39]:
X jTi j
SDR ¼ stdevðTÞ   stdevðTi Þ (5)
i
jTj
where stdev() is a standard deviation, j  j stands for the number of
instances, T is all instances, T1, T2,… are result sets from splitting
attributes. An attribute is determined as a node when it has a max-
imum SDR compared to all other attributes’ SDR. If no attribute
can reduce a standard deviation of class values, a model tree will
be identical to a simple linear regression model. For example,
there are two attributes and one class variable in the model tree
example in Fig. 5. stdev(T) is the standard deviation of the class
values. All possible split points of the two attributes are used to
estimate SDRs of the class values after splitting. Then, the one
with the maximum SDR becomes the split point and the attribute
of the split point becomes the node. The termination criterion of
splitting in the M5P is when the number of instances is less than
four or when the standard deviation at a node is less than
0.05*stdev(T). Once the splitting operation is finished, instances
at the leaf nodes are used to build linear models.
A pruning procedure can reduce size of a tree and the risk of
over-fitting. The M5P algorithm uses postpruning or backward
pruning, which means the pruning process starts after a tree
reaches a leaf node. If the lower estimated error is expected when
errors in nonleaf nodes and subtree are compared, the subtrees are
pruned to be leaves. The expected error of subtrees is the weighted
average of each node’s error by the proportion of sample sizes,
and the expected error of nonleaf nodes is [39]
X • Maximize unit profit of the product for its life cycle
jdeviation from predicted class valuej
n þ v instances
(6)
nv n • Uniqueness of design attributes
where n is the number of instances at the nonleaf node and v is the
number of parameters in a linear regression model in the node.
The second fraction represents the average of absolute difference
between the predicted value and the actual class value over each
of instances that reach the node. The first fraction is the compen-
sation factor to simplify the regression model in the node. In the
appendix, the manual implementation of the model tree in Fig. 5
is provided with sample data. Both splitting and pruning are con-
ducted based on the M5P algorithm.
The unique contribution of this Phase 1 is to propose a new
data generation scheme for a target time, which reflects two differ-
ent trends. By applying the model tree algorithm to this predicted
data set, this section shows some crucial properties that could not
be achieved with the previous models [16,2]: dynamic selection of
historical data, avoidance of over-fitting problem, identification of
performance information of constructed model, and allowance of
a numeric prediction. Section 3 will show empirical test results
with higher prediction accuracy.
2.2 Phase 2: Optimal product design for multiple profit
cycles. As shown in Fig. 1, products can have multiple life cycles
in the leasing market. When design engineers determine the initial
product design over the multiple life cycles, they should consider
not only the profit from the initial lease, but also the profit from
the recoveries and releases, which can be a hidden source of prof-
its. Usually, the latter part is ignored in the initial design stage due
to the absence of supporting models. The CPTM results from
Sec. 2.1 are expressed as model tree functions and will help to
reveal the hidden source of profits.
The optimal product design for multiple profit cycles is formu-
lated as a mathematical model and the overall architecture of the
model is depicted in Fig. 6. Model tree functions are used to
reflect customer preferences and technological obsolescence over
061002-6 / Vol. 136, JUNE 2014
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms
Fig. 6 Architecture of optimal design with CPTM
time. In order to address the reliability of target products over
time, a reusability function is formulated, which will give proba-
bilities of reusable and nonreusable products. The probabilities
affect the cost of end-of-life processes. While the optimizer evalu-
ates the unit profit of a given set of attributes that are decision var-
iables, model tree and reusability functions will take those
decision variables and return the unit price and the cost of end-of-
life process at a given time t, respectively.
2.2.1 Problem Statement. The unit profit of a design over
multiple life cycles is obtained by a mathematical model. The
model is summarized as the following optimization problem:
Objective
Constraints
Decision variables
• Target product design attributes
Given inputs
• Historical transactional data as a set of attributes and paid
price
• Generational information of parts
• Reliability information
• Cost of manufacturing and new parts
• Cost of reconditioning and logistics
2.2.2 Mathematical Formulation. Objective Function: The
objective function is expressed as the summation of unit profits,
which is the difference between unit price and unit cost at a given
time t as
X
1
Maximize f ¼ ðpt  ct Þ (7)
t ð1 þ rÞd
Since the multiple life cycles occur in the future, an annual inter-
est rate r should be applied to discount the value. For the present
value, ð1=ð1 þ rÞd Þ is multiplied and d is the number of the years.
A unit price at time t is derived from the model tree function
MTt() in Eq. (8). Binary decision variables, Yij, represent the level
of noncomponent based attributes such as weight, size, color, etc.
and Xij represent the level of component based attributes or re-
placeable and upgradable attributes such as battery, memory,
CPU, etc. A unit cost is divided into three different costs. First,
when time t is the starting time (t1), it is the production of new
products, and the unit cost consists of manufacturing costs and
forward logistics costs in Eq. (9). The manufacturing cost is
affected by Xij. If product attribute i has the level of j, Xij equals 1;
otherwise, it equals 0. Second, when time t is the take-back time,
it is the remanufacturing of take-back products, and the unit cost
Transactions of the ASME
 X
Table 3 Probability of reusable and non-reusable parts at dif- cEOLðtÞ ¼ cmanufacturing
j Xij ð1  bÞ þ creconditioning b (15)
ferent time t j

Time t¼1 t¼2  t¼n

The contribution of this Phase 2 is to formulate the optimal
2
Prob. of nonreusable parts (1b) (1– b) þ b(1b) ¼ (1b)    (1b) product design model with multiple life cycles. In order to address
Prob. of reusable parts b b2 þ (1b)b ¼ b  b some issues on the multiple life cycles such as customer preferen-
ces, technological obsolescence, and reliability over time, model
trees from the CPTM and reusability functions are combined in
the optimization model.
consists of end-of-life process costs, and inverse and forward lo-
gistics costs in Eq. (10). Third, when a product eventually reaches 3 Performance Test of CPTM
the point that is not profitable (tend), it will have a unit price of In this section, a set of different data are tested with the CPTM
recycling and a cost of disposal in Eqs. (11) and (12) algorithm. In order to understand the mechanism of the CPTM, Secs.
3.1 and 3.2. provide simple data sets. A real data set is also tested in
pt ¼ MT t ðYij ; Xij Þ; where t 6¼ tend (8) Sec. 3.3 to verify the performance of the CPTM algorithm in a real
X situation. Section 3.4 deals with the most complex data that will be
t
c ¼ cmanufacturing
j Xij þc forwardlogistics
; where t ¼ t1 (9) used for the statistical analysis and the application study in Sec. 4.
j
Four different static models were compared with the dynamic
ct ¼ cinverselogistics þ cEOLðtÞ þ cforwardlogistics ; where t ¼ ttake-back model, CPTM: linear regression, model tree (M5P), support vec-
tor machine (SMOreg), and neural network (Multilayer Percep-
(10) tron). Weka [40] was used to implement these models, and the
pt ¼ precycling ; where t ¼ tend (11) names in the parenthesis represent the equivalent algorithms. For
t inverselogistics forwardlogistics disposal
the automatic time series forecasting, R [41] was used with the
c ¼c þc þc ; where t ¼ tend package, forecast [37]. All static models construct a predictive
(12) model based on the latest data set (latest in Table 5) or all histori-
cal data sets (all in Table 5) as heuristics. On the other hand, the
Constraints: Equation (13) imposes that each product attribute CPTM utilizes all historical data selectively and builds a predic-
i has a unique attribute level j. In other words, finding a unique tive model based on the generated data set. It is important to real-
combination of each design attribute is the design problem. ize that the CPTM algorithm also uses the model tree but the
X X difference is in the use of trend embedded target data.
Yij ¼ 1; Yij 2 ð0; 1Þ; Xij ¼ 1; Xij 2 ð0; 1Þ (13) As a performance measure, mean absolute error (MAE) and
j j root mean-squared error (RMSE) were used [28]. Equation (16)
and (17) show the MAE and the RMSE with the predicted class
A probability of reusable parts b or a reusability function is values, b1 ; b2 ; …; bm and the actual class values, d1 ; d2 ; …; dm .
defined as the multiplication of each part’s reliability at time t in
Eq. (14). Equation (15) formulates the cost of end-of-life proc- jb1  d1 j þ    þ jbm  dm j
Mean Absolute Error ¼ (16)
esses as manufacturing costs with new parts and reconditioning m
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
costs with old parts. Reconditioning is conducted with probability
b. If a part is not reusable, then a new part should be used. jb1  d1 j2 þ    þ jbm  dm j2
Root Mean Squared Error ¼
Because a part’s manufacturing cost differs by design decisions, m
remanufacturing with new parts is formulated as a function of Xij (17)
with a probability of nonreusable parts (1b). Table 3 shows the
probability of reusable and nonreusable parts at different time t 3.1 Test With Data Generated From Stationary Linear
with the assumption that the reliability of a product will be back Mapping Function. The data shown in Fig. 7 were generated by
in a state as new after end-of-life processes. Therefore, memory- stationary linear mapping functions over time. There are two nom-
lessness is satisfied. inal attributes, A and B, and one class variable, Class. The values
YX of column A increase by two and those of column B by one
b¼ cj ðtÞXij (14) over time, which represents the first type of trend. In order to
i j generate the class values, the first five instances used a mapping

Fig. 7 Data from stationary linear mapping function and generated future data

Journal of Mechanical Design JUNE 2014, Vol. 136 / 061002-7

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 Table 4 Forecast results trend. Since this is a stationary case, all data sets from t ¼ 1 to
t ¼ 8 have the same mapping functions.
t¼7 t¼8 The goal is to construct a predictive model for t ¼ 8 data with
(Latest) Forecast (Target) t ¼ 1 to t ¼ 7 data sets. First, the values of each attribute and class
variable were normalized within a single time step. Second, based
A Min 13 15 15
Stationary Max 22 24 24
on Eq. (4), the smoothing factor, a ¼ 0, was selected using the
Linear B Min 11 12 12 built model tree from t ¼ 1 to t ¼ 6 data with different as on
Data Max 56 57 57 Eq. (3) and tested with t ¼ 7 data in terms of the MAE. Then, the
Class Min 11.3 12.42 12.4 selected a gave the number of samples from each normalized data
Max 29.7 30.73 30.5 set based on Eq. (3). Since a ¼ 0, all 70 normalized data were
sampled. Third, the automatic time series forecasting was con-
A Min 13 15 15 ducted for the original values of attributes A,B, and class variable
Stationary Max 22 24 24 Class in Table 4. By applying the predicted minimum and maxi-
nonlinear B Min 11 12 12 mum values to the sampled normalized data, Fig. 7 shows the
Data Max 56 57 57
resulted trend embedded data set for the target time t ¼ 8. Finally,
Class Min 11.75 12.37 11.79
Max 56.9 65.55 64.7 the model tree algorithm was applied to the generated data set and
the built model tree is the predictive model from the CPTM algo-
Real data Class Min 1 0.9 0 rithm. The model tree was pruned so that 10 linear models were
Max 379 379 395 reduced to only two linear models. The pruned model showed
almost the same performance accuracy compared to the unpruned
tree. Table 5 shows the result of the performance test.
Table 5 Performance results
3.2 Test With Data Generated From Stationary Nonlinear
MAE RMSE Mapping Function. The data shown in Fig. 8 were generated by
stationary nonlinear mapping functions over time. Two nominal
Dynamic CPTM 1.30 1.57
attributes, A and B are the same as in Sec. 3.1, but nonlinear map-
Stationary Linear Regression 3.93/3.83 5.12/5.06 ping functions were used: Class ¼ 0:01  A2 þ 0:9  B þ Random pffiffiffi
linear Static Model Tree 3.87/2.59 5.44/4.93 ð0:2 0:2Þ for the first five instances and Class ¼ 0:2  Aþ
Data (latest/all) SVM 3.80/3.71 5.12/5.31 0:3  B þ Randomð0:2 0:2Þ for the last five instances.
Neural Network 3.61/5.46 4.99/6.89 The goal is to construct a predictive model for t ¼ 8 data with
t ¼ 1 to t ¼ 7 data sets, and Fig. 8 shows the trend embedded data
Dynamic CPTM 6.77 8.88 set for the target time. The smoothing factor, a ¼ 0.3, was selected
Stationary Linear Regression 11.80/14.64 15.75/17.35 using the built model from t ¼ 1 to t ¼ 6 data and tested with t ¼ 7
nonlinear Static Model Tree 11.86/8.65 18.34/15.96 data. Also the automatic time series forecasting was conducted in
data (latest/all) SVM 12.73/15.68 18.19/20.98 Table 4. Based on the smoothing factor, 27 normalized instances
Neural Network 13.24/15.50 19.59/20.58
were sampled and the predicted values were applied to them. The
Dynamic CPTM 13.70 18.40 model tree was pruned so that eight linear models were reduced to
Linear Regression 21.9/25.13 31.1/33.54 only two linear models. The pruned model showed a little bit
Real data Static Model Tree 18.20/14.56 25.8/9.02 higher performance accuracy compared to the unpruned tree.
(latest/all) SVM 18.79/20.65 34.86/33.10 Table 5 shows the result of the performance test.
Neural Network 17.32/20.96 21.81/24.97
3.3 Test With Real Data. Second-hand values or buy-back
prices of cell phones [42] were tested with the CPTM. Since the
function, Class ¼ 0:1  A þ 0:9  B þ Randomð0:2 0:2Þ and data set was obtained with the list of target cell phones, all attrib-
the remaining five instances used a mapping function, ute values were the same but buy-back prices were varied over
Class ¼ 0:4  A þ 0:2  B þ Randomð0:2 0:2Þ with some ran- time. Due to market penetration, the market value of the same
domness in the functions, which represents the second type of products has a tendency to go down over time. After

Table 6 Example of data set (decision variables and snapshot of data)

Display Weight Hard CPU Graphics card Memory Battery Touchscreen Price
size (inch) (lbs) drive (GB) (technology) (technology) (GB) (hours) (technology) ($)

9 0.8 40 Core 2 duo HD G 4 6 Touch D p1

(Y 11) (Y 21) (X11) (X21) (X31) (X41) (X51) (X61)
10 1 80 Core 2 e HD G 2000 6 12 Touch C p2
(Y 12) (Y 22) (X12) (X22) (X32) (X42) (X52) (X62)
11 1.5 120 Core i3 HD G 2500 8 18 Touch B p3
(Y 13) (Y 23) (X13) (X23) (X33) (X43) (X53) (X63)
12 2 250 Core 2 i5 HD G 3000 16 24 Touch A ⯗
(Y 14) (Y 24) (X14) (X24) (X34) (X44) (X54) (X64)
320 Core 2 i7 HD G 4000 32
(X15) (X25) (X35) (X45)
500 Core 2 i7 e
(X16) (X26)

10 1.3 40 Core 2 i7 HD G 2500 4 6 Touch D 950

10.5 0.8 80 Core 2 duo HD G 2000 8 24 Touch B 910
12 0.9 320 Core 2 i5 HD G 4000 32 18 Touch C 1,200
⯗ ⯗ ⯗ ⯗ ⯗ ⯗ ⯗ ⯗ ⯗

061002-8 / Vol. 136, JUNE 2014 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 Fig. 8 Data from stationary nonlinear mapping function and generated future data
preprocessing the original data, monthly data sets of 155 cell
phones from June 2009 to March 2010 were tested with 10 differ-
ent attributes: camera pixel, talk time, touch screen, weight, mem-
ory slot, wiFi, MP3, GPS, bluetooth, and 3G.
The goal is to construct a predictive model for the t ¼ 10 data
with t ¼ 1 to t ¼ 9 data sets. The smoothing factor, a ¼ 0, was
selected using the built model from t ¼ 1 to t ¼ 8 data and tested
with t ¼ 9 data. Also the automatic time series forecasting was
conducted in Table 4. Table 5 shows the result of performance
test with this real data.
3.4 Test With Data Generated From Nonstationary Linear
Mapping Function. Twenty four data sets were generated ran-
domly with assumed ranges of attributes and some trends reflect-
ing real-world tablet PC leasing markets. Each data set has 200
instances and 8 different attributes shown in Table 6. The first part
of Table 6 explains levels of each attribute which are decision var-
iables explored in Sec. 2.2.2. The second part indicates an exam-
ple of the generated data. The data set shows transactional history
and the class variable is the price that customers paid. Since this is
data from nonstationary linear mapping functions, mapping func-
tions with some randomness vary over time.
The goal is to construct a series of predictive models for the
t ¼ n þ 1 data using t ¼ 1 to t ¼ n data sets. n were increased by 1
from 11 to 23. For example, for an unseen t ¼ 12 data set, static
models were constructed using the latest data set t ¼ 11 while the
CPTM mined a trend from t ¼ 1 to t ¼ 11 data sets and constructed
a tree model from a predicted data set at t0 ¼ 12 with the calcu-
lated smoothing factor 0.5. Then, both models were evaluated
Fig. 9 Comparison of one time-ahead prediction accuracy
between static and dynamic model
Journal of Mechanical Design
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms
with real t ¼ 12 data set in terms of the MAE. This procedure con-
tinued up to t ¼ 24 (total 13 times) for one time-ahead prediction
and the results are shown in Fig. 9.
In order to obtain a statistically valid conclusion between static
models and the CPTM, both parametric (F and T-test) and non-
parametric (Mann-Whitney) tests were employed. With a signifi-
cance level of alpha ¼ 0.05, the accuracy of the CPTM model was
significantly higher than that of static models with the generation
of trend embedded data.
3.5 Discussion. The CPTM algorithm showed good predic-
tive performances in comparison to the four well-known static
models in Table 5. From the cases of data sets generated from
simple stationary linear and nonlinear mapping functions, it is rel-
atively clear to look at the effect of the geometric sampling and
the time series prediction of attributes and class variables as
shown in Figs. 7, 8, and Table 4. The geometric sampling helped
to reflect the trend of relation between attributes and class varia-
bles over time. The automatic time series forecasting also gave
good approximations of future attribute and class values. In both
cases, smoothing factors were close to zero, which makes sense in
that stationary mapping functions were applied over time.
The real data in Sec. 3.3 had an interesting data structure. The
values of attributes were fixed but the class variable continued to
change, which is why there are only predictions for class variables
in Table 4. It is important to realize that even though the forecast
result was similar to the latest data, the geometric sampling
improved the predictive performance alone. Empirical tests showed
that without a precise prediction of attribute values, the CPTM
algorithm worked well with the geometric sampling. Also, without
any knowledge of customer preferences and their trend over time in
the used product market, the selected smoothing factor can indicate
that underlying relations between attributes and class variables
were quite stationary in the interval of one month. This work with
real data has great potential and can provide some directions from
data collection to real application in different design domains.
The last case of data from nonstationary mapping functions rep-
resents a very complex data structure, and the question was
whether the CPTM worked well in this case. Due to the nonsta-
tionary nature of data, the prediction error of the CPTM was close
to other static models, e.g., at t ¼ 13 and t ¼ 14, etc., in Fig. 9.
However, statistical results showed that an overall performance of
the CPTM is better than other models with this data.
Among those four static models, the model tree was selected for
the purpose of direct comparison since the CPTM algorithm also
adopts the model tree for the prediction of class variables. From all
the tested data cases, the predictive performance of the CPTM out-
performed that of the model tree, and this indicates that the genera-
tion of a trend embedded data set improved the accuracy.
While conducting these experiments, a total of five possible
sources of variation on the result were observed: smoothing
JUNE 2014, Vol. 136 / 061002-9
 factor, model tree, time series prediction, selection of samples or
random number generator, and size of samples. With the data sets
from simple stationary linear and nonlinear mapping functions, it
is difficult to construct a good model due to the small number of
instances. Random sampling can also have a great impact with the
small sample size. However, the impact from the last two factors
can be minimized with large-scale data. The model tree algorithm
in the CPTM is known to be fast or capable of dealing with large
number of instances and attributes [43]. Empirical tests with a
data set which is similar to the real buy-back price data (10 attrib-
utes) in Sec. 3.3 showed that the model tree took 1.34 s with 104
instances and 10.44 s with 105 instances running on an Intel Core
i5 2.5 GHz Processor.
Moreover, the important observations are the facts that the
model trees were pruned to avoid an overfitting and could gener-
ate performance information of the models by applying the 10-
fold cross-validation technique. Also, continuous attributes and
class variables were allowed in the models. These are aforemen-
tioned benefits of the CPTM over the DPTM from generating the
trend embedded future data.
4 Application
The overall methodology in Sec. 2 was applied to tablet PC design
in the leasing market. The same data sets described in Sec.3.4. were
used. Weka and R in Sec. 3 also provided necessary tools for the
model tree induction and the automatic time series forecasting.
4.1 Problem Setting. Tablet PCs are wireless, portable touch
screen-operated computers. It is assumed that feasible candidate
design attributes are defined in Table 6. It is expected that the start
of leasing time is t ¼ 12 and the company has accumulated data
sets from t ¼ 1 to t ¼ 11. A manufacturer (and lessor) should man-
age multiple life cycles of its tablet PC by taking back leased
products and releasing after processing for the next usage-life.
The goal of this problem is to find the optimal tablet designs for
multiple profit cycles while considering customer preferences,
technological obsolescence, and reliability. Given inputs and
assumptions are as follows:
Given inputs
• Historical transactional data as a set of attributes and price
• Generational difference information in the appendix
• Reliability information in the appendix
• Manufacturing and new parts cost in the appendix
• Reconditioning cost ¼ $120
• Logistics cost: forward logistics ¼ $5, reverse logistics ¼ $5
Assumptions
• 2-year time frame (No disposal stage)
• Leasing period is fixed at six months
• After end-of-life processes, the reliability of a product will be
back in a state as new
• Upgrade is not considered so that there are no compatibility
issues during EOL processes
• Time for logistics and remanufacturing is negligible compared
to that of the leasing period length
4.2 Applying CPTM. Since it was assumed that the tablet
PC will have 6 months leasing time over the 2-year time frame,
the number of life cycles is 4 and predictive models from t ¼ 12 to
t ¼ 15 are needed by design. Static models were constructed using
heuristics e.g. only the t ¼ 11 data set or all historical data. For the
CPTM, it predicted 1 time, 2 time, 3 time, and 4 time-ahead data
sets using t ¼ 1 to t ¼ 11 data sets selectively and built model trees
from the predicted data sets. Table 7 presents the results. At t ¼ 12
all split points of the 8 attributes were used to estimate SDRs of
the class vales after splitting. Since the split point 2.5 of the attrib-
ute CPU maximized the standard deviation reduction of the class
values, the CPU became the first node with branches of less than
061002-10 / Vol. 136, JUNE 2014
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms
Table 7 CPTM results of illustration example
At t ¼ 12; MT 12 is defined as follows:
CPU  2:5: LM1
CPU > 2.5:
j Hard drive  2:5: LM2
j Hard drive > 2.5: LM3
LM num: 1
Class(Price) ¼  1.0472 * Weight þ 0.6497 * Hard Drive  8.8437
* CPU  28.1698 * Graphics Card 17.0256 * Memory  10.779
* Battery life  8.6369 * Touchscreen þ 1014.3514
LM num: 2
Class(Price) ¼ 7.8489 * Display size  3.4677 * Weight  12.11
* Hard drive  16.2778 * CPU  30.4275 * Graphics card  20.6762
* Memory  3.8257 * Battery life  1.9893 * Touchscreen þ 940.6956
LM num: 3
Class(Price) ¼ 8.6802 * Display size  3.4677 * Weight  4.4711
* Hard Drive  18.8712 * CPU  21.8863 * Graphics card  2.8183
* Memory  32.5704 * Battery life  1.9893 * Touchscreen þ 990.9373
At t ¼ 13; MT 13 is defined as follows:
Hard drive  2:5: LM1
Hard drive > 2.5: LM2
LM num: 1
Class(Price) ¼ 8.3753 * Display size  33.9229 * Hard Drive  20.2652
* CPU  37.4899 * Graphics card  11.9472 * Memory  7.3387
* Battery life  8.3119 * Touchscreen þ 993.0399
LM num: 2
Class(Price) ¼ 0.7584 * Display size  21.0345 * Weight  6.1539
* Hard Drive  18.1878 * CPU  8.9427 * Graphics card  7.4079
* Memory  31.1933 * Battery life  0.505 * Touchscreen þ 1103.8625
At t ¼ 14; MT 14 is defined as follows:
LM1
LM num: 1
Class(Price) ¼ 16.3777 * Display size þ 10.698 * Hard drive  17.8321
* CPU  20.8245 * Graphics card  14.2723 * Memory  19.7954
* Battery life þ 853.3004
At t ¼ 15, MT15 is defined as follows:
LM1
LM num: 1
Class(Price) ¼ 7.481 * Hard drive  27.5058 * CPU  14.6723
* Graphics card  18.1991 * Memory  28.7345 * Battery life  6.7278
* Touchscreen þ 1107.0882
or equal to 2.5 and greater than 2.5. After all of the splitting,
instances at the leaf nodes were used to build linear models. The
resulted model tree was pruned so that it had only three linear
models, while the original tree had 169 linear models. By pruning
the tree, the built model’s performance was decreased based on the
generated data or training data (e.g., 14.6% more errors than the
unpruned tree) but the prediction accuracy of the model was
improved with real t ¼ 12 data (e.g., 1.3% less errors than the
unpruned tree) due to the generalization. The 10-fold cross-validation
in Weka was used to get the performance information of the0 built
model from the training data (i.e., the predicted data set at t ¼ 12)
and the prediction accuracy was calculated from the real data (i.e.,
the real data set at t ¼ 12). This shows that unpruned trees have a
strong chance to be over-fitted. The comparison result between static
models with the latest data set and the CPTM is shown in Fig. 10.
Finally, these linear models will be used in the optimization model.
4.3 Design for Multiple Profit Cycles. Table 8 shows the
mathematical formulation of the application derived from Sec.
2.2.2. The objective function consists of unit profits from four
lease contracts and the interest rate, 3%, was assumed. Prices or
market values that reflect the trend of customer preferences and
Transactions of the ASME

Fig. 10 Comparison of 1, 2, 3, and 4 time-ahead prediction ac-
curacy between static and dynamic model
technological obsolescence were formulated with the model tree
functions depicted in Table 7.
4.4 Discussion. Similar to the previous section for the CPTM
performance, Fig. 10 indicates that the accuracy of the CPTM
outperformed that of those static models even with the multiple
time-ahead predictions. The accuracy was measured by the mean
absolute error which is the average deviation between predicted
and observed class variable price. Based on Fig. 10, the CPTM
was adopted as the predictive model for this application.
The CPTM result in the Table 7 shows model trees constructed
from the CPTM algorithm. At t ¼ 12 and t ¼ 13, multiple linear
regression models were built and at t ¼ 14 and t ¼ 15, simple
regression models were formulated to explain the class variable
price. At t ¼ 12, the model tree consists of three linear models:
LM1, LM2, and LM3. When the attribute CPU has a generational
difference less than 2.5, the first linear model, LM1, is selected. If
the attribute CPU has a generational difference greater than 2.5
then the attribute hard drive will work as a splitting criterion.
Again, if the attribute hard drive has a generational difference less
than 2.5, then the second linear model, LM2, will be selected.
Otherwise, the third linear model, LM3, will be used. In each lin-
ear model, eight different design attributes in Table 6 with a con-
stant term explain the class variable.
Excel solver with an evolutionary algorithm was used to solve
the illustration design problem. The selected designs are shown in
Table 9, and the total life cycle unit profits are revealed in Table
10. The selected best design attributes are 12-in. in display size,
0.8-lbs in weight, 40-GB in hard drive, Core 2 i7 e in CPU, HD G
4000 in graphics card, 8-GB in memory, 24-hour in battery life,
and touch C in Touch screen with the total life cycle unit profit of

Table 8 Mathematical formulation for illustration example

Objective function
1 1 1
Maximize f ¼ ðp12  c12 Þ þ ðp13  c13 Þ þ ðp14  c14 Þ þ ðp15  c15 Þ
ð1:03Þ0:5 ð1:03Þ ð1:03Þ1:5
p12 ¼ MT 12 ðYij ; Xij Þ, p13 ¼ MT 13 ðYij ; Xij Þ; p14 ¼ MT 14 ðYij ; Xij Þ; p15 ¼ MT 15 ðYij ; Xij Þ
X manufacturing
c12 ¼ cj Xij þ cforwardlogistics
j
c13 ¼ cinverselogistics þ cEOLð13Þ þ cforwardlogistics
c14 ¼ cinverselogistics þ cEOLð14Þ þ cforwardlogistics
c15 ¼ cinverselogistics þ cEOLð15Þ þ cforwardlogistics
Constraints
X
h1 : Yij ¼ 1
j
X
h2 : Xij ¼ 1
j
X manufacturing
h3 : cEOLðtÞ ¼ cj Xij ð1  bÞ þ creconditioning b
Q X j 
h4 : b ¼ cj Xij
i j
h5 : Yij ; Xij 2 ð0; 1Þ

Table 9 Result of optimal tablet pc design

Display size Weight Hard drive CPU Graphics card Memory Battery Touch screen
(inch) (lbs) (GB) (technology) (technology) (GB) (hours) (technology)

CPTM 12 0.8 40 Core 2 i7 e HD G 4000 8 24 Touch C

Linear Regression (latest/all) 12 / 12 0.8 / 0.8 40 / 40 Core 2 duo/Core 2 i7 e HD G 4000/HD G 3000 32/32 24/24 Touch D/Touch A

Table 10 Result of total life cycle unit profit

t ¼ 12 t ¼ 13 t ¼ 14 t ¼ 15 Total life cycle

CPTM Profit($) 419 430 386 327 1562

Price($) 949 994 951 875 3769
Cost($) 530 564 565 548 2207
Linear Regression(latest/all) Profit($) 477/392 390/418 346/376 232/343 1445/1529
Price($) 962/972 919/1000 875/958 745/909 3501/3839
Cost($) 485/580 529/582 529/582 513/566 2056/2310

Journal of Mechanical Design JUNE 2014, Vol. 136 / 061002-11

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 $1,562. There are other design results from linear regression mod- Nomenclature
els (i.e., static model) with the two heuristics in Table 9. First, the a¼ number of instances for each data set
“only latest data set case” selected different CPU, memory, and at ¼ number of instances that will be sampled at time
touch screen attributes. It is interesting that the model generated CPTM ¼ continuous preference trend mining
much more profits at t ¼ 12 but the design selected by the CPTM cdisposal ¼ unit cost of disposal of a take-back product
brought more profits over the life cycle as shown in Table 10. Sec- cEOL(t) ¼ unit cost of end-of-life processes at time t
ond, the “all data set case” selected different graphic card, mem- cforwardlogistics ¼ unit cost of forward logistics from factory to
ory, and touch screen attributes. This model generated more customers
profits than the other heuristic but fewer profits than the CPTM. cmanufacturing ¼ unit cost of manufacturing a product
The illustration concludes that the proposed framework can iden- creverselogistics ¼ unit cost of reverse logistics from customers to
tify the optimal design that maximizes the total life cycle profit factory
based on historical transactional data sets. ct ¼ unit cost of a product at time t
From the result obtained from the artificially generated data, it EOL ¼ end-of-life
can be argued that customers are very sensitive about the techno- i¼ index for attributes
logical obsolescence for CPU, graphics card and battery attributes j¼ index for levels or options under one attribute
(refer to Table 6). Manufacturers should use the latest cutting MAE ¼ mean absolute error
edge technology for these parts. At the same time, due to the pop- MTt ¼ model tree at time t
ularity of cloud storage and external storage devices, the capacity precycling
¼ unit price of recycling a take-back product
of a hard drive seems to have become less important to customers. pt ¼ unit price of a product at time t
This suggests that manufacturers can place less priority on hard r¼ interest rate
drive capacity. The proposed framework enabled this type of RMSE ¼ root mean squared error
insight, which is not readily available under the previous trend SDR ¼ standard deviation reduction
mining approaches. Xij ¼ component based attribute as a binary decision
The illustration does not consider the option to upgrade for the variable
initial design selection problem. However, an additional decision Yij ¼ noncomponent based attribute as a binary deci-
making process can determine the proper end-of-life options sion variable
including upgrades [35]. Given the target time, manufacturers can a¼ smoothing factor or characteristic of product
decide whether the decrease of generational difference (i.e., domain
upgrade) is better than reconditioning for each component. The b¼ probability of reusable parts
life cycle management plan can then be set up including upgrades. cj ¼ reliability of reusable part j

5 Conclusion and Future Work

Appendix
In this paper, a new predictive trend mining algorithm, CPTM,
is developed in the context of product and design analytics. Unlike
traditional, static data mining algorithms, the CPTM does not
assume stationarity, and dynamically extracts valuable knowledge A.1 Manual Implementation of Model Tree. Based on the
of customers over time. By generating trend embedded future sample data in Table 11, the model tree in Fig. 5 is manually built.
data, the CPTM algorithm not only shows higher prediction accu- The sample data has two attributes, A and B, and one class
racy in comparison with static models but also provides essential variable C.
properties that could not be achieved with the previous trend min-
ing algorithms: dynamic selection of historical data, avoidance of
over-fitting problem, identification of performance information of Table 11 Sample data for model tree
constructed model, and allowance of a numeric prediction. Also,
A B C
the optimization model for multiple life cycles is formulated as a
binary integer programming model and combined with the CPTM 200 14.5 10
result. Using the proposed framework, design engineers can select 140 20 26
the optimal design for the target product that can generate multi- 90 14.4 29
ple profit cycles. The illustration example of tablet PC design 98 13.5 32
showed that the optimization model with the CPTM can reveal 86 16 34
hidden profit cycles successfully. 50 24 44
In the future, it will be interesting to observe the impact of the
prediction interval in the CPTM algorithm even though there are Table 12 Determining a root node of model tree
multiple sources of variation as discussed in Sec. 3.5. Different
optimal design solutions can be obtained based on the interval. C A midpoint stdev(T) stdev(T1) stdev(T2) SDR
The optimization model in the illustration example was simplified
in order to show the application of the CPTM. Additionally, com- 44 50 68 11.2 N/A 9.5 N/A
patibilities among different parts, different product life cycles, 34 86 88 11.2 7.1 9.8 2.3
and product families can be considered for more interesting and 29 90 94 11.2 7.6 11.4 1.7
realistic problems. Finally, instead of having a set of attributes as 32 98 119 11.2 6.5 11.3 3.1
a priori, capturing of emerging attributes and management of 26 140 170 11.2 6.9 N/A N/A
10 200 N/A
dynamic attribute sets would be possible tasks in the future.
C B midpoint stdev(T) stdev(T1) stdev(T2) SDR

Acknowledgment 32 13.5 14 11.2 N/A 12.4 N/A

29 14.4 14.5 11.2 2.1 14.4 0.9
The work presented in this paper was supported by the National 10 14.5 15.3 11.2 11.9 9.0 0.7
Science Foundation under Award No. CMMI-0953021. Any opin- 34 16 18 11.2 11 12.7 0.4
ions, findings and conclusions or recommendations expressed in 26 20 22 11.2 9.5 N/A N/A
this publication are those of the authors and do not necessarily 44 24 N/A
reflect the views of the National Science Foundation.

061002-12 / Vol. 136, JUNE 2014 Transactions of the ASME

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 Table 13 Determining the second node of model tree Table 15 Assumed information of reliability

C A midpoint stdev(T) stdev(T1) stdev(T2) SDR t ¼ 11 Hard drive CPU Graphics card Memory Battery Touch screen

44 50 68 6.5 N/A 2.5 N/A 1 1 1 1 1 1

34 86 88 6.5 7.1 2.1 1.9 1 1 1 1 1 1
29 90 94 6.5 7.6 N/A N/A 1 1 1 1 1 1
32 98 N/A 1 1 1 1 1 1
1 1 1 1
C B midpoint Stdev (T) Stdev (T1) stdev(T2) SDR 1 1
32 13.5 14 6.5 N/A 7.6 N/A t ¼ t þ 1 Hard drive CPU Graphics card Memory Battery Touch screen
29 14.4 15.2 6.5 2.1 7.1 1.9
34 16 20 6.5 2.5 N/A N/A 0.95 0.95 0.96 0.98 0.99 0.99
44 24 N/A 0.98 0.95 0.97 0.98 0.99 0.97
0.99 0.98 0.98 0.99 0.99 0.97
0.99 0.99 0.98 0.99 0.98 0.91
0.99 0.99 0.93 0.96
Table 12 shows all standard deviation reduction (SDR) calcula- 0.95 0.93
tions for determining the root node and splitting point of the
model tree. First, the class variable and each attribute are grouped,
and values of the attribute are sorted from smallest to largest. For
each mid-point, calculate the standard deviation of class values in When the value of attribute A is greater than 119, only two
the divided groups. For example, with the mid-point 88 (the sec- instances reach the node. The termination criterion of the M5P
ond row), stdev(T1) and stdev(T2) represent the standard devia- algorithm (i.e., less than four instances) stops further splitting for
tion of {44, 34} and {29, 32, 26, 10}. stdev(T) represents the this branch. For the other branch, there are four instances and the
standard deviation of all the class values. Then, the final column standard deviation at the node (6.5) is greater than the other crite-
SDR is calculated based on Eq. (5), and the midpoint that rion (i.e., 0.05*stdev(T) ¼ 0.56). After removing the instances that
produces the maximum SDR is the splitting point (i.e., 119 of are greater than 119, the same procedure can be applied as shown
attribute A). in Table 13. In this case, two splitting points (i.e., 88 of attribute
A and 15.2 of attribute B) produce the same SDR so that either of
Table 14 Assumed information of generational difference them can be selected and the model performance will be the same.
Figure 5 shows the case that 15.2 of attribute B is selected. All the
t ¼ 11 Hard drive CPU Graphics card Memory Battery Touch screen nodes have less than four instances so that the splitting operation
of the model tree is completed. Finally, the instances at the leaf
5 5 4 4 3 3 nodes are used for regression models. Due to the small number of
4 4 3 3 2 2 instances, all leaf nodes take a simple model i.e. LM1: C ¼ 18,
3 3 2 2 1 1 LM2: C ¼ 30.5, and LM3: C ¼ 39.
2 2 1 1 0 0 A pruning procedure compares the expected error of leaf nodes
1 1 0 0
and a nonleaf node. The nonleaf node B has two leaf nodes
0 0
and their expected error can be calculated as follows: the absolute
t ¼ 12 Hard drive CPU Graphics card Memory Battery Touch screen difference between the predicted and the actual class value is
averaged at the each node and weighted by the proportion of
6 5 4 5 3 3 sample sizes (ð2=4Þðjð29  30:5j=2Þ þ ðj32  30:5j=2ÞÞ þ ð2=4Þ
5 4 3 4 2 2 ððj34  39j=2Þ þ ðj44  39j=2ÞÞ ¼ 3:25Þ. The internal regression
4 3 2 3 1 1 model at the node B (C ¼ 1.31*B þ 12.59) is then used to calcu-
3 2 1 2 0 0
2 1 0 1
late the expected error (2.05) based on Eq. (6). Also by dropping
1 0 the parameter of the internal regression model (C ¼ 34.75),
another expected error (4.63) can be calculated but this model can
t ¼ 13 Hard drive CPU Graphics card Memory Battery Touch screen be ignored due to the higher expected error. Since the expected
error of the node B is lower than that of the leaf nodes, the tree
6 6 5 5 5 3 should be pruned and the internal regression model becomes a
5 5 4 4 4 2
leaf node. Similarly, by comparing the node A and leaf nodes, the
4 4 3 3 3 1
3 3 2 2 2 0 pruning operation can be determined and it turns out that the tree
2 2 1 1 should be pruned. After the pruning procedure, one regression
1 1 model (C ¼ 0.21*A þ 52.21) replaces the three regression
models.
t ¼ 14 Hard drive CPU Graphics card Memory Battery Touch screen

8 7 6 6 5 5 A.2 Assumed Information in Application. Tables 14–16

7 5 5 5 4 4 show the assumed information of generational difference, reliabil-
6 4 4 4 3 3 ity, and cost for manufacturing and new parts, respectively.
5 3 3 3 2 2
4 2 2 2 Table 16 Assumed information of cost for manufacturing and
3 1 new parts ($)

t ¼ 15 Hard drive CPU Graphics card Memory Battery Touch screen Hard drive CPU Graphics card Memory Battery Touch screen

9 8 6 7 5 5 40 60 90 20 40 50
8 7 5 6 2 4 55 70 110 30 50 60
7 6 4 5 1 3 75 90 130 40 60 85
6 4 3 4 0 2 90 100 140 80 70 100
5 3 2 3 100 110 160 135
4 2 120 125

Journal of Mechanical Design JUNE 2014, Vol. 136 / 061002-13

Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms

 References
[1] Tucker, C. S., and Kim, H. M., 2008, “Optimal Product Portfolio Formulation
by Merging Predictive Data Mining With Multilevel Optimization,” ASME J.
Mech. Des., 130(4), pp. 991–1000.
[2] Tucker, C. S., and Kim, H. M., 2011, “Trend Mining for Predictive Product
Design,” ASME J. Mech. Des., 133(11), p. 111008.
[3] Van Horn, D., Olewnik, A., and Lewis, K., 2012, “Design Analytics: Capturing,
Understanding and Meeting Customer Needs Using Big Data,” ASME
International Design Engineering Technical Conferences and Computers and
Information in Engineering Conference (IDETC/CIE2011), Paper No.
DETC2012-71038.
[4] Tucker, C. S., 2011, “Data Trend Mining Design for Predictive Systems
Design,” Ph.D. thesis, University of Illinois, Chicago, IL.
[5] Rai, R., 2012, “Identifying Key Product Attributes and Their Importance Levels
From Online Customer Reviews,” ASME International Design Engineering
Technical Conferences and Computers and Information in Engineering Confer-
ence (IDETC/CIE2011), Paper No. DETC2012-70493.
[6] Environmental Protection Agency, 2011, “Electronics Waste Management in the
United States Through 2009,” U.S. EPA, May, Report EPA No. 530-R-11-002.
[7] Sodhi, M. S., and Reimer, B., 2001, “Models for Recycling Electronics End-Of-
Life Products,” OR Spektrum, 23(1), pp. 97–115.
[8] Fishbein, B. K., 1998, “EPR: What Does it Mean? Where is it headed?,” P2:
Pollution Prevention Rev., 8(4), pp. 43–55.
[9] Product Stewardship Institute, 2012, “Extended Producer Responsibility State
Laws.” Available at: http://productstewardship.us (accessed in May 2013).
[10] Wagner, S., 2003, Understanding Green Consumer Behaviour: A Qualitative Cog-
nitive Approach, Consumer Research and Policy Series, Taylor & Francis Group.
[11] Environmental Protection Agency, 2011, “Benefits of the Remanufacturing
Exclusion: Background Document in Support of the Definition of Solid Waste
Rule,” June, Washington, DC.
[12] Hucal, M., 2008, “Product Recycling Creates Multiple Lives for Caterpillar
Machines,” Peoria Magazines, September.
[13] King, A., Miemczyk, J., and Bufton, D., 2006, “Photocopier Remanufacturing
at Xerox uk a Description of the Process and Consideration of Future Policy
Issues,” Innovation in Life Cycle Engineering and Sustainable Development, D.
Brissaud, S. Tichkiewitch, and P. Zwolinski, eds., Springer Netherlands, pp.
173–186.
[14] Parker, D., and Butler, P., 2007, “An Introduction to Remanufacturing.” Avail-
able at: http://www.remanufacturing.org.uk (accessed in May 2013).
[15] Kusiak, A., and Smith, M., 2007, “Data Mining in Design of Products and Pro-
duction Systems,” Annu. Rev. Control, 31(1), pp. 147–156.
[16] B€ottcher, M., Spott, M., and Kruse, R., 2008, “Predicting Future Decision Trees
From Evolving Data,” Proceedings of ICDM’08, pp. 33–42.
[17] B€ottcher, M., 2011, “Contrast and Change Mining,” Wiley Interdiscip. Rev.:
Data Mining Knowledge Discovery, 1(3), pp. 215–230.
[18] Klinkenberg, R., 2004. “Learning Drifting Concepts: Example Selection vs.
Example Weighting,” Intell. Data Anal., 8(3), pp. 281–300. Available at: http://
www.iospress.nl/
[19] Ma, J., Kwak, M., and Kim, H. M., 2014. “Demand Trend Mining for Predictive
Life Cycle Design,” J. Clean. Prod.
[20] Vapnik, V. N., 1998, Statistical Learning Theory, Wiley-Interscience,
Hoboken, NJ.
[21] Fixson, S. K., 2004, “Assessing Product Architecture Costing: Product life
cycles, Allocation Rules, and Cost Models,” ASME International Design Engi-
neering Technical Conferences and Computers and Information in Engineering
Conference (IDETC/CIE2004), Paper No. DETC2004-57458.
[22] Duverlie, P., and Castelain, J. M., 1999, “Cost Estimation During Design Step:
Parametric Method Versus Case Based Reasoning Method,” Int. J. Adv. Manuf.
Technol., 15(12), pp. 895–906.
061002-14 / Vol. 136, JUNE 2014
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 04/14/2014 Terms of Use: http://asme.org/terms
[23] Seo, K., Park, J., Jang, D., and Wallace, D., 2002, “Approximate Estimation of
the Product Life Cycle Cost Using Artificial Neural Networks in Conceptual
Design,” Int. J. Adv. Manuf. Technol., 19(6), pp. 461–471.
[24] Zhao, Y., Pandey, V., Kim, H. M., and Thurston, D., 2010, “Varying Lifecycle
Lengths Within a Product Take-Back Portfolio,” ASME J. Mech. Des., 132(9),
p. 091012.
[25] Hyndman, R., Koehler, A., Ord, J. K., and Snyder, R., 2008, Forecasting with
Exponential Smoothing: The State Space Approach, Springer-Verlag, Berlin,
Heidelberg.
[26] Quinlan, J. R., 1993, C4.5: Programs for Machine Learning, Morgan Kauf-
mann Series in Machine Learning, Morgan Kaufmann Publishers.
[27] Quinlan, J. R., 1986, “Induction of Decision Trees,” Mach. Learn., 1(1), pp.
81–106.
[28] Witten, I., and Frank, E., 2005, Data Mining: Practical Machine Learning
Tools and Techniques, 2nd ed., The Morgan Kaufmann Series in Data Manage-
ment Systems, Elsevier Science.
[29] Cheung, K.-W., Kwok, J. T., Law, M. H., and Tsui, K.-C., 2003, “Mining Cus-
tomer Product Ratings for Personalized Marketing,” Decision Support Syst.,
35(2), pp. 231–243.
[30] Archak, N., Ghose, A., and Ipeirotis, P. G., 2011, “Deriving the Pricing Power
of Product Features by Mining Consumer Reviews,” Manage. Sci., 57(8), pp.
1485–1509.
[31] Ferreira, L., Jakob, N., and Gurevych, I., 2008, “A Comparative Study of Fea-
ture Extraction Algorithms in Customer Reviews,” 2008 IEEE International
Conference on Semantic Computing, pp. 144–151.
[32] Abulaish, M., Jahiruddin, Doja, M. N., and Ahmad, T., 2009, “Feature and
Opinion Mining for Customer Review Summarization,” Proceedings of the 3rd
International Conference on Pattern Recognition and Machine Intelligence,
PReMI’09, Springer-Verlag, pp. 219–224.
[33] Decker, R., and Trusov, M., 2010, “Estimating Aggregate Consumer
Preferences From Online Product Reviews,” Int. J. Res. Market., 27(4), pp.
293–307.
[34] De’ath, G., 2002, “Multivariate Regression Trees: A New Technique for Mod-
eling Species-Environment Relationships,” Ecology, 83(4), pp. 1105–1117.
[35] Kwak, M., and Kim, H. M., 2011, “Market-Driven Positioning of Remanufac-
tured Product for Design for Remanufacturing With Part Upgrade,” ASME
International Design Engineering Technical Conferences and Computers and
Information in Engineering Conference (IDETC/CIE2011), Paper No.
DETC2011-48432.
[36] Yue, S., Pilon, P., and Cavadias, G., 2002, “Power of the Mannkendall and
Spearman’s Rho Tests for Detecting Monotonic Trends in Hydrological
Series,” J. Hydrol., 259(14), pp. 254–271.
[37] Hyndman, R., and Khandakar, Y., 2008, “Automatic Time Series Forecasting:
The Forecast Package for R,” J. Stat. Softw., 27(3), pp. 1–22.
[38] Quinlan, J. R., 1992, Learning With Continuous Classes, World Scientific, Sin-
gapore, pp. 343–348.
[39] Wang, Y., and Witten, I. H., 1997, “Inducing Model Trees for Continuous
Classes,” Proceedings of the 9th European Conference on Machine Learning
Poster Papers, pp. 128–137.
[40] Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., and Witten, I.
H., 2009, “The Weka Data Mining Software: An Update,” SIGKDD Explor.
Newsl., 11(1), pp. 10–18.
[41] R Development Core Team, 2008, R: A Language and Environment for Statisti-
cal Computing, R Foundation for Statistical Computing, Vienna, Austria.
[42] Kwak, M., Kim, H. M., and Thurston, D., 2012, “Formulating Second-Hand
Market Value as a Function of Product Specifications, Age, and Conditions,”
ASME J. Mech. Des., 134(3), p. 032001.
[43] Shrestha, D. L., and Solomatine, D. P., 2006, “Machine Learning Approaches
for Estimation of Prediction Interval for the Model Output,” Neural Netw.,
19(2), pp. 225–235.
Transactions of the ASME